Statistical challenges in single-cell RNA-seq analysis III

July 6, 2019

Preview

Goal: identify a set of genes with high variability attributable to biology (over and above technical variability)
Useful for vizualization, dimension reduction, clustering, marker gene selection, etc
Challenging without orthogonal measurement of technical variability (e.g. spikeins)
- with spikeins: select genes with variance significantly above mean-variance trend in control genes
- without spikeins: select genes with variance significantly above overall mean-variance trend in all endogenous genes (assumes variance of most genes is purely technical)

Useful to summarize & visualize relationships between cells in low dimensional space
Commonly used approaches:
- PCA (Principal Components Analysis)
- tSNE (t-distributed Stochastic Neighbor Embedding)
- UMAP (Uniform Manifold Approximation and Projection)
Clustering can be carried out on reduced dimensions, but with caution

PCA attempts to extract the largest components of variation in the data
Nonlinear methods such as tSNE and UMAP attempt to map points to a global coordinate system that preserves local structure
- density & distance of points not preserved
- better at visualizing rare subtypes

Traditional differential expression for bulk RNA-seq aims at detecting shifts in mean (fold change)
Typically read counts are modeled with a two parameter distribution (mean, dispersion)